The symptom is usually described as a variety problem: “all my sounds sound the same.” The fix that gets tried first is more variants — more files, more randomisation, more pitch offsets. That fix often makes the problem worse, because the sameness was not caused by too few files. It was caused by every file being pulled toward the same set of parameter values.
A sound designer working alone tends to converge on a small set of settings that “sound good” through the same monitors, with the same reference tracks, for the same kind of event. A UI click that is clear and pleasant gets a sharp attack, a short decay, and a bright midrange. A pickup that is rewarding gets the same sharp attack, a slightly longer decay, and a bright midrange. A hit that is satisfying gets the same shape again. By the time there are forty sounds, they have forty different waveforms and one shared identity.
This is a parameter convergence problem, and it has to be fixed at the source — in the pitch centre, the spectral contour, the temporal envelope, and the frequency band each sound is allowed to own. Adding variants on top of converged parameters produces more versions of the same sound, not more distinct sounds.
Why More Variants Do Not Fix Sameness
The standard anti-repetition advice is correct for repetition: if the same event fires repeatedly, a random container with several takes and a pitch offset stops the ear from flagging the exact repeat. Wwise’s Randomizer applies a random offset to a property on every playback, and the Random Container picks from a pool with options to avoid the most recent selection[reference:0][reference:1]. That is the right tool for “this sword swing sounds identical every time.”
It is the wrong tool for “this sword swing sounds like the coin pickup.” A random container varies the same sound; it does not change what the sound is. If the sword and the coin share a pitch centre, an attack shape, and a spectral tilt, randomising each of them independently produces two families of similar sounds, and the player still cannot tell them apart at speed.
The distinction matters because the two problems have different fixes. Repetition is solved at the runtime layer with randomisation and pooling. Sameness is solved at the asset layer by making the parameters of each sound orthogonal to the parameters of the other sounds in its category. The sound variation system guide covers the runtime layer; this article is about the layer underneath it.
Convergence Point 1: Every Sound Gets the Same “Good” Envelope
The most common convergence point is the amplitude envelope. A short attack, a quick decay to a low sustain, and a clean release is the shape that makes almost any sound read as “snappy” and “professional” in isolation. It is also the shape that makes a UI click, a coin, a hit, and a jump all feel like the same event with different pitch.
The envelope is not a polish parameter. It is one of the primary carriers of what kind of event a sound is. A sound with a near-instant attack and a 40 ms decay reads as a contact; a sound with a 30 ms attack and a 400 ms decay reads as a release; a sound with a slow swell reads as a build. If every sound in a set has the contact shape, the set reads as forty contacts.
The fix is not to make the envelopes worse. It is to assign each event a temporal role before choosing the envelope, and to let the role dictate the attack and decay rather than letting the same “snappy” preset dictate both. The SFX duration guide gives the category ranges that make an event read as itself; the sameness fix is to use those ranges as a constraint on the envelope, not as a target to optimise toward.
Convergence Point 2: Everything Sits in the Same Midrange
The second convergence point is spectral. The midrange is where small speakers are most efficient, where dialogue intelligibility lives, and where a sound is easiest to make “clear.” A designer checking sounds on a laptop or a pair of earbuds will naturally push every sound toward the same 500 Hz–2.5 kHz region, because that is the region the device reproduces best.
The result is a set where every sound is clear and no sound is distinct. The auditory system separates simultaneous sounds by frequency band and by timbral contour; when every sound occupies the same band, the only remaining separation is temporal, and a fast game does not always provide it.
The fix is a frequency identity per category, assigned before the sound is made. A commonly used architecture for action audio assigns sub energy below 80 Hz to physical impact, low-mid 200–500 Hz to weight and body, high-mid 2–6 kHz to sharpness and attack, and air 8–16 kHz to surface texture and distance[reference:2]. A gunshot built from a 200–800 Hz body layer, a 2–8 kHz crack, and a sub layer below 100 Hz is distinct from a footstep not because it is louder, but because it owns different bands[reference:3].
The layering guide covers how to split layers so they do not compete. The sameness problem is the same principle applied across categories: if two events share a band, one of them has to move, or they will fuse into a single perceived texture regardless of how different the waveforms are.
| Band | What it carries | Which events should own it |
|---|---|---|
| Below 80 Hz | Physical impact, pressure | Explosions, heavy impacts, large weapons |
| 200–500 Hz | Weight and body | Footsteps, body hits, wooden surfaces |
| 2–6 kHz | Sharpness, attack | Gunshots, metal clicks, UI confirms |
| 8–16 kHz | Surface texture, air | Cloth, gravel, distant reflections |
Convergence Point 3: Every Sound Uses the Same Pitch Centre and the Same Pitch Contour
The third convergence point is pitch. Pitch is the most obvious way to distinguish two sounds, and it is also the easiest parameter to adjust, which means it gets adjusted constantly without being assigned. A set where every sound sits near the same fundamental, with the same small upward or downward slide, reads as one sound played at different speeds.
Pitch has two separable roles that get conflated. The pitch centre is the sound’s identity — where it lives in the register. The pitch contour is what the pitch does over the sound’s duration — rising, falling, flat, two-stage. A coin and a power-up can share a rising contour and still be distinct if their centres are far apart; a coin and a hit can share a centre and still be distinct if their contours move in opposite directions.
The fix is to assign both, separately. A practical approach is to give each event category a pitch register (low, mid, high) and a contour direction (rise, fall, flat), and to check that no two categories share both. The blip frequency tuning guide covers the register side for UI sounds; the same logic applies to any short effect. A semitone is the natural step for a rising or falling sequence, and a pitch multiplier of 2^(1/12) ≈ 1.06 shifts by one semitone[reference:4], which is a useful unit for deciding how far apart two registers need to be.
Convergence Point 4: Every Sound Has the Same Spectral Shape
The fourth convergence point is subtler than the first three and more pervasive. Two sounds can have different pitch centres and different envelopes and still feel like the same sound if their spectral contour is the same — if the energy is distributed across the spectrum in the same way, with the same tilt and the same movement over time.
Timbre research identifies a small set of spectral features that listeners use to order sounds: spectral centroid, spectral spread, spectral skewness, spectral slope, and the odd-to-even harmonic ratio[reference:5]. These are not abstract descriptors. They correspond to audible properties: centroid to brightness, spread to the width of the spectral energy, skewness to whether the energy leans low or high, slope to the overall tilt, and odd-to-even ratio to hollow versus full character.
A set where every sound has the same tilt and the same centroid trajectory will read as one family even when the fundamental frequencies differ. The fix is to assign a spectral character per category — bright and narrow for UI, dark and wide for impacts, tilted low for weight, tilted high for texture — and to check that the assigned character is actually present in the rendered sound rather than only in the intention.
The reference matching guide covers how to extract these features from a reference without copying it. The sameness use of the same method is different: you are not matching a reference, you are checking that your own sounds do not all match each other.
A Source-Level Separation Test
The test below is designed to be run on the raw assets, before any runtime randomisation is applied. If the set fails it, no amount of variation at the container level will fix the sameness, because the sameness is in the sources.
- Play every sound in a category in random order, without labels. Ask someone who did not make them to group the sounds by “what happened.” If the groupings do not match the intended categories, the sounds are not separable at the source.
- Play the same set with pitch randomised by ±1 semitone. If the randomisation makes the set more distinguishable, the identity was carried by pitch alone and the other parameters are converged. If it makes the set less distinguishable, the randomisation is crossing the identity threshold and needs to be narrowed.
- Look at the spectrum of each sound at the attack and at the tail. If every sound has the same centroid trajectory — the same rise, the same fall, the same flat sustain — the spectral contour is converged even if the waveforms differ.
- Check the envelope attack and decay times as numbers, not by ear. If every sound in the set has an attack under 10 ms and a decay under 100 ms, the set is almost certainly converged on the contact shape, regardless of how different the sources sound.
A set that passes this test has sounds that are distinct before any randomisation is applied. A set that fails it has sounds that are variations of one identity, and the fix belongs at the synthesis stage: change the pitch register, change the envelope role, change the frequency band, change the spectral tilt — one at a time, until the sounds are separable in a blind grouping.
The reason to fix this at the source rather than at the container is that runtime variation and source identity are different kinds of difference. A random container makes the same sound less predictable; it does not make it a different sound. If the sword swing and the coin pickup are built from the same envelope, the same band, and the same pitch contour, pooling them with three variants each produces two families of the same idea. The player will not hear repetition. They will hear a game where everything sounds like everything else, and no amount of randomisation will change that, because the sameness was never in the playback. It was in the parameters.